JPH10252537A - Control device for internal combustion engine - Google Patents

Abstract

PROBLEM TO BE SOLVED: To prevent a control value from becoming a remarkably improper value, so that good engine operation control can be performed, while utilizing a characteristic of a neural net. SOLUTION: A neural net inputting a plurality of engine operation parameters including a detection value of an air/fuel ratio sensor is used, cylinder-wise equivalent ratio HATAF is calculated (S1), a change amount ΔA/F thereof is calculated (S2). In accordance with the cylinder-wise equivalent ratio change amount ΔA/F, a basic fuel amount Ti is corrected, a fuel injection amount Tout is calculated (S4). A value of the engine operation parameter input to the neural net is limited within a range narrower than a range of parameter value which can be actually taken during operation of an engine.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a control device for controlling the operation of an internal combustion engine by controlling a fuel supply amount, an ignition timing and the like of the internal combustion engine using a neural network.

[0002]

2. Description of the Related Art An air-fuel ratio control device that controls the fuel supply amount of an internal combustion engine using a neural network has already been proposed (for example, Japanese Patent Application Laid-Open No. Hei 8-74627). In this device, a detected air-fuel ratio, a fuel injection amount, an engine speed, an intake pipe internal pressure, and the like are used as input parameters of a neural network, and a fuel supply amount is determined according to an output value of the neural network.

[0003]

However, in the above-described conventional apparatus, the detected parameters or the calculated parameters are directly input to the neural network, so that the input parameter values are used in the learning for determining the coupling coefficients of the neural network. In such a case, the output value of the neural network may be extremely inappropriate. Further, in order to avoid such a situation, it is necessary to perform learning using all the parameter values of the operable range of the engine, but since the operable range of the engine is very wide, in an actual use state, It is very difficult to learn all operating parameters, and the time required for the learning test is enormous.

In a control device for an internal combustion engine, it has been conventional to monitor a calculated control amount such as a fuel injection amount and limit the control amount so that the control amount falls within a range of predetermined upper and lower limits. Is being done. However, in the case of the neural network, if the input parameter value exceeds the assumed range (learning range), the output value becomes indefinite and the reliability of the output value is lost, so that the output value, that is, the control amount is limited. Even when the processing was performed, there were cases where the value was in the limit range and was an inappropriate value. That is, in the limit processing for limiting the control amount within the range of the upper and lower limits, it is sometimes impossible to avoid a situation where the control amount becomes inappropriate.

The present invention has been made in view of this point, and it is possible to prevent the control amount from becoming extremely inappropriate and to perform good engine operation control while utilizing the characteristics of the neural network. It is an object of the present invention to provide a control device for an internal combustion engine that can be used.

[0006]

In order to achieve the above object, the present invention relates to a control apparatus for an internal combustion engine, which controls the operation of the internal combustion engine by using a neural network which inputs operating parameters of the internal combustion engine. An input limiting unit is provided for limiting the value of the operating parameter input to the net to a range narrower than a range of values that can be actually taken during the operation of the engine.

[0007] According to this configuration, the value of the engine operation parameter input to the neural network is limited to a range narrower than the range of values that can be actually taken during operation of the engine, and the operation of the engine is controlled.

[0008]

Embodiments of the present invention will be described below with reference to the drawings.

FIG. 1 is an overall configuration diagram of an internal combustion engine (hereinafter simply referred to as “engine”) and an air-fuel ratio control device thereof according to an embodiment of the present invention.
In the middle of the intake pipe 2, a throttle valve 3 is provided. A throttle valve opening (θTH) sensor 4 is connected to the throttle valve 3, outputs an electric signal corresponding to the opening of the throttle valve 3, and supplies it to an electronic control unit (hereinafter referred to as “ECU”) 5. .

A fuel injection valve 6 is provided for each cylinder between the engine 1 and the throttle valve 3 and slightly upstream of an intake valve (not shown) of the intake pipe 2, and each injection valve is connected to a fuel pump (not shown). The fuel injection time (valve opening time) is controlled by a signal from the ECU 5 while being electrically connected to the ECU 5.

On the other hand, an intake pipe absolute pressure (PBA) sensor 12 is provided immediately downstream of the throttle valve 3, and an absolute pressure signal converted into an electric signal by the absolute pressure sensor 12 is supplied to the ECU 5. . Further, an intake air temperature (TA) sensor 13 is attached downstream thereof, detects the intake air temperature TA, outputs a corresponding electric signal, and supplies the electric signal to the ECU 5.

The engine water temperature (TW) sensor 14 mounted on the main body of the engine 1 is composed of a thermistor or the like, detects the engine water temperature (cooling water temperature) TW, outputs a corresponding temperature signal, and supplies it to the ECU 5. Engine speed (NE)
The sensor 15 and the cylinder discrimination (CYL) sensor 16 are mounted around a camshaft (not shown) of the engine 1 or around a crankshaft. The engine speed sensor 15 is the engine 1
A pulse (hereinafter referred to as a "TDC signal pulse") at a predetermined crank angle position every time the crankshaft rotates by 180 degrees, and the cylinder discriminating sensor 16 outputs a signal pulse at a predetermined crank angle position of a specific cylinder. These signal pulses are supplied to the ECU 5.

In FIG. 1, an exhaust pipe 21 is
The exhaust pipe 21 is provided with a catalytic converter (three-way catalyst) for purifying components such as HC, CO, and NOx in exhaust gas. ) 23 are arranged. Exhaust pipe 2
On the upstream side of the first catalytic converter 23, a proportional air-fuel ratio sensor 22 (hereinafter referred to as "LAF sensor 22") is mounted. The LAF sensor 22 detects the oxygen concentration in the exhaust gas, and detects the detected value. Output an electrical signal according to EC
Supply to U5.

Next, the exhaust gas recirculation mechanism (EGR) will be described.

An exhaust recirculation passage 25 is provided between the intake pipe 2 and the exhaust pipe 21 in a bypass shape. The exhaust gas recirculation path 2
5 is an exhaust pipe 2 whose one end is upstream of the LAF sensor 22.
1 and the other end is connected to the intake pipe 2.

An exhaust gas recirculation amount control valve (hereinafter, referred to as an EGR valve) 26 is interposed in the exhaust gas recirculation passage 25. The EGR valve 26 includes a valve chamber 27 and a diaphragm chamber 28.
A wedge-shaped valve body 30 disposed in the valve chamber 27 and movably arranged in the vertical direction so that the exhaust gas recirculation passage 25 can be opened and closed; and a valve shaft 31.
Diaphragm 32 connected to the valve body 20 through
And a spring 33 for urging the diaphragm 32 in the valve closing direction.
It is composed of In addition, the diaphragm chamber 28
Atmospheric pressure chamber 3 defined on the lower side via diaphragm 32
4 and a negative pressure chamber 35 defined on the upper side.

The atmosphere chamber 34 communicates with the atmosphere via a vent 34a, while the negative pressure chamber 35 is connected to a negative pressure communication passage 36. That is, the negative pressure communication passage 36 is connected to the intake pipe 2.
And the absolute pressure (negative pressure) P in the intake pipe in the intake pipe 2
BA is introduced into the negative pressure chamber 35 via the negative pressure communication passage 36. An atmosphere communication passage 37 is connected in the middle of the negative pressure communication passage 36, and a pressure regulating valve 38 is provided in the middle of the atmosphere communication passage 37. The pressure regulating valve 3
Reference numeral 8 denotes a normally closed solenoid valve, and the atmospheric pressure or the negative pressure is selectively supplied to the negative pressure chamber 35 of the diaphragm chamber 28 through the pressure regulating valve 38, and the negative pressure chamber 35 is controlled by a predetermined control. Generate pressure.

Further, the EGR valve 26 is provided with a valve opening (lift) sensor 39.
9 detects the operating position (valve lift amount LACT) of the valve element 30 of the EGR valve 26, and outputs the detection signal to the ECU 5
To supply. The EGR control is executed after the engine is warmed up (for example, when the engine coolant temperature TW is equal to or higher than a predetermined temperature).

The ECU 5 shapes input signal waveforms from various sensors, corrects a voltage level to a predetermined level, and converts an analog signal value into a digital signal value. The CPU 5b includes a storage unit 5c for storing various calculation programs executed by the CPU 5b, calculation results, and the like, an output circuit 5d for supplying a drive signal to the fuel injection valve 6, and the like.

The CPU 5b estimates an air-fuel ratio for each cylinder of the engine 1 and sets a target air-fuel ratio on the basis of the above-mentioned various engine operating parameters.
According to the estimated cylinder-by-cylinder air-fuel ratio and target air-fuel ratio, the TD
The fuel injection time T of the fuel injection valve 6 in synchronization with the C signal pulse
out is calculated. Since the fuel injection time Tout is proportional to the fuel injection amount by the fuel injection valve 6, it is also referred to as a fuel injection amount in this specification. The CPU 5b calculates a target lift amount LCMD of the EGR valve 26 based on various engine operation parameters, and controls the pressure adjustment valve 38 so that the detected valve lift amount LACT matches the target lift amount LCMD.

The CPU 5b outputs, via an output circuit 5d, drive signals for the fuel injection valve 6 and the pressure regulating valve 38 based on the results calculated as described above.

In this embodiment, the ECU 5 constitutes a cylinder-by-cylinder air-fuel ratio estimating means, an air-fuel ratio estimating means, an air-fuel ratio controlling means, a first calculating means and a second calculating means.

The above-described estimation of the cylinder-by-cylinder air-fuel ratio by the CPU 5b is performed using a neural network that inputs various engine operating parameters, and calculates the cylinder-by-cylinder equivalent ratio HATAF by converting the air-fuel ratio into an equivalent ratio. FIG. 2 is a diagram for explaining a schematic structure of the neural network employed in the present embodiment. This neural network has a three-layer structure of an input layer, an intermediate layer, and an output layer, and its learning algorithm is based on a well-known back-propagation (Back-Pro
pagation) Learning algorithm was adopted. Note that the learning algorithm may employ another method such as a random search method.

As shown in FIG. 2, information input to each cell (neuron) in the input layer is an engine operation parameter Xi (i = 1 to n), and the information is weighted by a coupling coefficient matrix. Is input to each cell in the middle layer. In the intermediate layer, the output is determined for each cell by, for example, a sigmoid function, and similarly to the processing from the input layer to the intermediate layer, the output weighted by the coupling coefficient matrix is input to the output layer, and the cylinder equivalent ratio ATAF Is output. Each coefficient of the coupling coefficient matrix is actually installed with an air-fuel ratio sensor at the exhaust port of each cylinder to actually detect the air-fuel ratio, and for each operation parameter value, the detected cylinder-specific air-fuel ratio as teacher data, It is determined by the back propagation learning algorithm so that the total error function is minimized.

In this embodiment, three neural nets having the structure shown in FIG. 2 are provided. Specifically, engine 1
A neural network for starting when the engine is started (during cranking), and a neural network for when the LAF sensor 22 is inactive after the engine 1 is started and the LAF sensor 22 is not activated. A neural network for activation of the LAF sensor, which is used when the LAF sensor 22 is activated after the engine 1 is started, is provided. A coupling coefficient matrix corresponding to each neural network is stored in the storage means of the ECU 5. Have been.

As the engine operation parameters input to the input layer of the neural network for the activation of the LAF sensor, the following are employed in this embodiment with n = 13. Further, as the engine operation parameters input to the input layers of the neural network for the inactive state of the LAF sensor and for the starting time, X1 to X10 excluding the following X11 to X13 are adopted as n = 10 in the present embodiment. .

X1 = Tout (k) (fuel injection time (current value)) X2 = Tout (k-4) (fuel injection time (value before 4TDC)) X3 = TA (intake temperature) X4 = TW (engine water temperature) X5 = NE (engine speed) X6 = PBA (absolute pressure in intake pipe) X7 = θTH (throttle valve opening) X8 = LACT (EGR valve lift) X9 = KCMD (k) (target air-fuel ratio coefficient (current value) )) X10 = KCMD (k-4) (Target air-fuel ratio coefficient (4TDC
X11 = KACT (k) (Detection equivalent ratio (current value)) X12 = KACT (k-4) (Detection equivalent ratio (value before 4TDC)) X13 = ΔKACT (k) (Detection equivalent ratio change Quantity = KAC
T (k) -KACT (k-4)) Here, the target air-fuel ratio coefficient KCMD is a coefficient by which the basic fuel amount Ti is multiplied by a calculation formula of a fuel injection amount Tout described later, and depends on the engine operating state. The set target air-fuel ratio (A / F) is converted into an equivalent ratio. The detected equivalent ratio KACT is obtained by converting the air-fuel ratio detected by the LAF sensor 22 into an equivalent ratio. Subscript (k),
(k-4) is added to indicate that the value is the current value and the value 4 TDC before (the period before the TDC signal pulse is generated four times), and in this embodiment, the engine 1
Is a four-cylinder, for example, Tout (k) and Tout (k-
4) is the fuel injection time corresponding to one specific cylinder. All parameters for which the subscript (k) is omitted represent the current value.

In the following description, the input parameters are represented by X1 to Xn in order to avoid complexity of notation.

FIG. 3 is a flowchart of a process for calculating the fuel injection amount Tout. This process is executed in synchronization with the generation of the TDC signal pulse.

First, in step S1, the HAT shown in FIG.
By the AF calculation processing, the cylinder-by-cylinder equivalent ratio HATAF is calculated using the neural network.

In step S101 of FIG. 4, each engine operation parameter is input (read), and then normalization of the input data is performed (step S102).

Specifically, as shown in FIG. 5, first, a limit check of input data is performed (step S11).
1). That is, each input data is set to the input upper / lower limit values LMTINH, LMTINL preset for each input data.
And if the input data exceeds the input upper limit LMTINH, the input data is set to the input upper limit LMTINH, and the input lower limit LMTIH is set.
If it is lower than NL, a process for setting the input lower limit value LMTINL is performed. For example, the input lower limit value and the input upper limit value of the engine speed NE are respectively set to NELMTNILL, N
Assuming that ELMTINH is satisfied, if NE <NELMTINL, NE = NELMTINL, and NE> NEL
When it is MTINH, NE = NELMTINH. Here, input upper and lower limit values LMTINH, LMTINL
Is set to the maximum and minimum values of the data used in learning to determine the coupling coefficient matrix of the neural network,
Assuming that upper and lower limits that each input data can take when the device is normal are LMTH and LMTL, LMTL ≦ LMTI
It has a relationship of NL <LMTINH ≦ LMTH. That is, in the present embodiment, the value of the input data is limited to a range narrower than the range of values that can be actually taken during operation of the engine.

This is because the characteristics of the neural network, that is, the output value corresponding to the input data value exceeding the learning range of the neural network may deviate significantly from the desired value. By using in place of the closest value in the above, even if the desired value itself cannot be obtained as the output value of the neural network, it is possible to obtain an output value that does not significantly reduce the control performance. is there. By adopting such a configuration, it is possible to prevent the range of data used in learning from becoming too wide, and to reduce the number of steps required for determining a coupling coefficient matrix of a neural network. In addition, a problem such as a case where a limit process is performed on the output value does not occur, and the final control amount (fuel injection amount) becomes extremely inappropriate while taking advantage of the characteristics of the neural network. No good engine control can be performed.

The input upper and lower limits are set, for example, as follows (indicated by "input lower limit to input upper limit").

X1, X2 (TOUT): 2 msec to 8 msec X3 (TA): −25 ° C. to 50 ° C. X4 (TW): −25 ° C. to 85 ° C. X5 (NE): At start 200 rpm to 500 rpm After start 1000 rpm to 3000 rpm X6 (PBA): 260 mmHg to 560 mmHg X7 (θTH): 0 ° at start (fully closed fixed) 10 ° to 30 ° after start X8 (LACT): LACTMIN (fully closed) to LACTMAX (fully open) X9, X10 (KCMD) : 0.67 to 1.16 (A / F conversion 22 to 12.7) X11, X12 (KACT): (KCMD-0.136) to (KCMD + 0.135) (A / F conversion (A / F (KCMD) ) +2) to (A / F (KCMD) −2) X13 (ΔKACT): −0.034 to 0.034 (A / F conversion 0.5 to −) .5) where the input on the lower limit value of the detected equivalent ratio KACT, based on the target air-fuel ratio coefficient KCMD corresponding point, ± 2A
/ F.

In the following step S112, the input data X
i (i = 1 to n) is applied to the following equation to obtain normalized data N
Xi is calculated.

NXi = (Xi−CXi) / CXi (i = 1 to n) where CXi is the median of the input upper and lower limits.
In the input data, the detected equivalent ratio change amount ΔKA
Since the median value of CT is “0”, this normalization operation is not performed.

As described above, by normalizing, all the input data are converted into data having a median value of "0", so that the calculation of the sigmoid function is executed only by providing one table of the sigmoid function. It is possible to do.

Returning to FIG. 4, in step S103, FIG.
The calculation of the intermediate layer shown in FIG. That is, in step S121 of FIG. 6, it is determined whether or not a start flag FSTART indicating "1" indicating that the engine is being started is "1".
When FSTART = 1 and at the time of starting, a coupling coefficient matrix for starting is selected (step S125), and FS
If TART = 0 and after starting, it is determined whether or not an activation flag FLAFACT indicating “1” that the LAF sensor 22 is activated is “1” (step S122). And if FLAFACT = 0 and LA
When the F sensor 22 is inactive, a coupling coefficient matrix for when the LAF sensor is inactive is selected (step S124), and the FLAF is selected.
When ACT = 1 and the LAF sensor 22 is active, LA
A coupling coefficient matrix for F sensor activation is selected (step S123). At this time, in steps S123 to S125, not only the first coupling coefficient matrix [aji] used in the operation of the following step S125 but also the second coupling coefficient matrix [bj] used in the operation of the output layer described later are simultaneously performed. select. This selection processing corresponds to processing for switching between three neural nets.

Here, whether or not the LAF sensor 22 has been activated is determined by, for example, the resistance value of the sensor element or the LA value.
This is performed by a known method using a resistance value of a heater for hastening the activation of the F sensor 22.

In the following step S126, the first coupling coefficient matrix [aji] is added to the normalized data NXi obtained by the following equation (1).
(I = 1 to n, j = 1 to m) are multiplied by a matrix operation (step S126) to perform first intermediate variables YA1 to YAm. Here, the number m of the first intermediate variables YAj corresponds to the number of cells in the intermediate layer, and for example, m is about 20. When the number of cells is increased, the accuracy is improved but the amount of calculation is increased. Therefore, the number m of cells is determined in consideration of both.

[0042]

(Equation 1) Next, a sigmoid function operation by table search is performed on the first intermediate variables YA1 to YAm, and the second intermediate variable YB
1 to YBm are calculated (step S122). In this embodiment, Equation 2 is used as the sigmoid function, and the input / output characteristics of this function are as shown in FIG. Since Equation 2 is an odd function symmetrical with respect to the origin, a table for calculating y for the region of x ≧ 0 in FIG. 8 is actually set. For the region of x <0,
By performing a table search with | x | and setting the sign of the search value y to a minus value, a sigmoid function operation is performed.

[0043]

[Mathematical formula-see original document] y = 2 / (1 + exp (-2x))-1 Here, when Expression 2 is represented as y = SGM (x), the operation in step S127 is represented as Expression 3.

[0044]

## EQU3 ## YBj = SGM (YAj) (j = 1 to m) Returning to FIG. 4, in step S104, the operation of the output layer is performed. That is, as shown in FIG. 7, the second intermediate variable YBj
In the second coupling coefficient matrix (row vector) [bj] (j = 1
To m) to calculate a third intermediate variable YC (step S131), and a sigmoid function operation to perform a table search on the third intermediate variable YC and calculate output data Z (step S132). I do. These operations are represented by the following Expressions 4 and 5.

[0045]

(Equation 4)

[0046]

## EQU5 ## In step S105 of FIG. 4, the output data Z is inversely normalized, and the cylinder-by-cylinder equivalent ratio HATAF is calculated. The inverse normalization processing is performed by the following equation (6).

[0047]

HATAF = Z × HATAFC + HATAFC Here, HATAFC is an equivalent ratio of 1. that is equivalent to the stoichiometric air-fuel ratio.
Median value set to zero.

The calculations of the above equations 1 to 6 are executed in synchronization with the generation of the TDC signal pulse, so that the cylinder-by-cylinder equivalent ratios HATAF (k), HATAF (k + 1), HAT are sequentially output.
AF (k + 2) and HATAF (k + 3) represent equivalent ratios corresponding to each of the four cylinders, and the corresponding cylinder of HATAF (k + 4) is H
Same as ATAF (k).

The coupling coefficient matrices [aji] and [bj] (j = 1 to m) used in Equations 1 and 4 are determined in advance by learning using teacher data as described above. .

As described above, in this embodiment, the cylinder-by-cylinder equivalent ratio HATAF is calculated by using a neural network. Therefore, compared to a case where the cylinder-by-cylinder equivalent ratio is calculated by using an observer, the cylinder having no delay is used. The estimation of another air-fuel ratio can be performed. That is, an accurate cylinder-by-cylinder air-fuel ratio can be obtained even in a transient state in which the engine operating state changes.

The above is the cylinder-by-cylinder equivalence ratio calculation processing in step S1 of FIG. 3. In step S2 of FIG.
It is determined whether or not the start flag FSTART is "1", and FS is determined.
If TART = 0 and it is not at the time of starting, it is determined whether or not the activation flag FLAFACT is "1" (step S3). When FSTART = 1 and the engine is started, or when FLAFACT = 0 and the LAF sensor 22 is not activated, the process proceeds to step S4.

In step S4, time series data HATAF (k), HATAF (k + 4), H
A high-pass filter process is performed on ATAF (k + 8), HATAF (k + 12),. Specifically, the cylinder-by-cylinder equivalent ratio change amount ΔA / F is calculated by, for example, the following equation.

ΔA / F = HATAF (k) −HATAF (k−4) The high-pass filter processing is not limited to this. For example, a well-known averaging processing or smoothing processing is performed on the time-series data. Then, a process of subtracting the data after the averaging or the averaging process from the original data may be performed.

In the following step S5, a limit process for the cylinder-by-cylinder equivalent ratio change amount ΔA / F is performed. Specifically, for example, the cylinder-by-cylinder equivalent ratio change amount ΔA / F is equal to the change amount upper limit value ΔA / FL.
MTH (for example, 0.034 (A / F conversion -0.5))
Is exceeded, ΔA / F = ΔA / FLMTH, and the variation lower limit value ΔA / FLMTL (for example, −0.034 (A
/ F conversion 0.5)), ΔA / F = Δ
A / FLMTL.

Next, the fuel injection amount To is calculated by the following equation (7).
ut is calculated (step S6).

[0056]

## EQU7 ## Tout = Ti × KCMD + KP1 × ΔA / F
+ TIVB Here, Ti is a basic fuel amount set according to the engine speed NE and the intake pipe absolute pressure PBA.
D is the aforementioned target air-fuel ratio coefficient, KP1 is a proportional gain set to, for example, a negative predetermined value (for example, -0.05), TI
VB is a correction term that is set according to the output voltage of a battery that supplies drive power to the fuel injection valve 6.

The calculation based on Equation 7 is also equivalent to the cylinder equivalent ratio HAT.
Since it is executed in synchronization with the generation of the TDC signal pulse similarly to AF, the sequentially output fuel injection amounts Tout (k), Tout (k)
out (k + 1), Tout (k + 2), and Tout (k + 3) represent the fuel injection amount corresponding to each of the four cylinders, and Tout (k
The corresponding cylinder of +4) is the same as Tout (k).

In the processing of steps S4 to S6 (corresponding to the first calculating means described in the claims), before the activation of the LAF sensor 22, the detected equivalent ratio KACT is not used as an input parameter of the neural network. Considering that the reliability of the estimated absolute value of the cylinder-specific equivalence ratio HATAF becomes low due to the individual difference of the engine or the effect of purging the fuel vapor into the intake pipe, the variation ΔA of the cylinder-specific equivalence ratio HATAF is considered.
The feedback control according to only / F is performed.

On the other hand, in step S3, FLAFACT =
When the LAF sensor 22 is activated, the cylinder-by-cylinder air-fuel ratio correction coefficient KAF # N (N = 1 to 4) is calculated by PID control according to the cylinder-by-cylinder equivalent ratio HATAF (step S7). ). Specifically, the following formula 8
Is applied to the deviation DKAF (= KCMD-HATAF) between the target equivalent ratio KCMD and the cylinder equivalent ratio HATAF.
A proportional term KAFP, an integral term KAFI, and a derivative term KAFD are calculated, and a cylinder-specific correction coefficient KAF # N is calculated by Expression 9.

[0060]

KAFP (k) = DKAF (k) × KP2 KAFI (k) = DKAF (k) × KI2 + KAFI (k−4) KAFD (k) = (DKAF (k) −DKAF (k−4)) × KD
2 Here, KP2, KI2, and KD2 are a proportional gain, an integral gain, and a derivative gain, respectively.

[0061]

KAF # N = KAFP (k) + KAFI (k) + KAFD (k) In the next step S8, the fuel injection amount Tout is calculated by the following equation (10).

[0062]

Tout = Ti × KCMD × KAF # N + TIVB The processing in steps S7 and S8 (corresponding to the second calculating means described in the claims) is performed after the activation of the LAF sensor 22 and the detected equivalent ratio KACT. Since the cylinder-by-cylinder equivalence ratio HATAF is calculated by using a neural network having the input parameter as a parameter, feedback control is performed by ordinary cylinder-by-cylinder PID control in consideration of the high reliability of the absolute value of the HATAF value. It was done.

Using the equation (7) or (10), the fuel injection amount To
By calculating ut, it is possible to perform control so that the air-fuel ratio of each cylinder matches the air-fuel ratio corresponding to the target air-fuel ratio coefficient KCMD according to the cylinder equivalent ratio HATAF calculated using the neural network, It is possible to perform air-fuel ratio feedback control that is superior in responsiveness and stability as compared with control using a conventional observer.

Further, three neural nets for calculating the cylinder-by-cylinder equivalent ratio HATAF are provided for starting, when the LAF sensor is inactive, and when the LAF sensor is active, and the operating state of the engine or the active state of the LAF sensor is provided. In this case, an appropriate cylinder-by-cylinder equivalent ratio can be obtained in each state. Then, since the fuel injection amount Tout is controlled according to the cylinder-by-cylinder equivalence ratio HATAF thus obtained, more accurate air-fuel ratio control can be performed over a wide range of the engine operating state.

When the LAF sensor is inactive, feedback control is performed in accordance with the variation ΔA / F of the cylinder-by-cylinder equivalent ratio HATAF.
Since the normal cylinder-by-cylinder feedback control according to the deviation DKAF between the ATAF and the target air-fuel ratio coefficient KCMD is performed, appropriate cylinder-by-cylinder air-fuel ratio control is performed according to the reliability of the absolute value of the cylinder equivalent ratio HATAF. be able to.

Note that the present invention is not limited to the above-described embodiment, and various modifications are possible. For example,
Regarding the neural network for starting and the time when the LAF sensor is inactive, the input data may be the following operating parameters.

X1 = Tout (k) (fuel injection time (current value)) X2 = Tout (k-4) (fuel injection time (value before 4 TDC)) X3 = Tout (k-12) (fuel injection time ( X4 = TW (engine water temperature) X5 = NE (engine speed) X6 = PBA (intake pipe absolute pressure) X7 = TCOUNT (T from engine start time)
In the above-described embodiment, the plurality of neural nets are used for starting, when the LAF sensor is inactive, and when the LAF sensor is active. However, after the engine is started, the engine water temperature TW is changed. Depending on whether the temperature is higher or lower than a predetermined temperature (for example, 30 ° C.), the neural network for high temperature and the neural network for low temperature may be switched and used. Neural networks for the activation and inactivation of the LAF sensor (a total of four neural nets) may be provided when the engine water temperature TW is high and low, respectively. Further, a plurality of operating regions defined by the engine speed NE and the intake pipe absolute pressure PBA may be set, and a neural network corresponding to each of the plurality of operating regions may be provided.

The cylinder equivalent ratio HATAF (k), HA
TAF (k + 1), HATAF (k + 2), HATAF (k + 3), H
The time series data of ATAF (k + 4),... Are appropriately weighted and averaged to calculate an assembly equivalent ratio HATAFMIX in the exhaust pipe assembly, and the fuel injection is performed in accordance with the assembly equivalent ratio HATAFMIX. The quantity Tout may be calculated. Alternatively, the neural network itself may be configured to output the aggregate equivalent ratio HATAFMIX.

In the above-described embodiment, the process of limiting the input parameter values of the neural network is performed by controlling the ignition timing of the engine, controlling the opening amount of the EGR valve, or controlling the intake valve by using the neural network inputting the engine operating parameters. Alternatively, the present invention can be applied to control of the valve timing of an exhaust valve.

[0070]

According to the present invention, as described in detail above, the value of the engine operation parameter input to the neural network is:
Since the operation of the engine is controlled within a range narrower than the range of values that can be actually taken during operation of the engine, even if the engine operation parameter value becomes extremely large or small, the output of the neural network is not changed. Good engine operation control can be performed without remarkably deviating from the desired value.
Also, since the range of the input data of the neural network can be narrowed, the number of steps for determining the coupling coefficient of the neural network by learning can be reduced.

[Brief description of the drawings]

FIG. 1 is a diagram showing a configuration of an internal combustion engine and a control device thereof according to an embodiment of the present invention.

FIG. 2 is a diagram showing a configuration of a neural network.

FIG. 3 is a flowchart of a fuel injection amount calculation process.

FIG. 4 is a flowchart of a process for calculating a cylinder equivalent ratio.

FIG. 5 is a flowchart of processing for normalizing input data.

FIG. 6 is a flowchart of a process corresponding to an operation in a hidden layer of the neural network.

FIG. 7 is a flowchart of a process corresponding to a calculation in an output layer of the neural network.

FIG. 8 is a diagram for explaining a sigmoid function.

[Explanation of symbols]

 REFERENCE SIGNS LIST 1 internal combustion engine 2 intake pipe 4 throttle valve opening sensor 5 electronic control unit (input limiting means) 6 fuel injection valve 12 intake pipe absolute pressure sensor 13 intake temperature sensor 14 engine water temperature sensor 15 engine speed sensor 21 exhaust pipe 22 air-fuel ratio Sensor

Claims (1)

[Claims] 1. A control device for an internal combustion engine that controls the operation of an internal combustion engine by using a neural network that inputs an operation parameter of the internal combustion engine, wherein a value of the operation parameter input to the neural network is determined by the engine An input limiting means for limiting the value to a range narrower than a range of values that can be actually taken during the operation of the internal combustion engine.